kata-research-phase
Fail
Audited by Gen Agent Trust Hub on Apr 10, 2026
Risk Level: HIGHCOMMAND_EXECUTIONPROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill performs shell operations in its process steps that use the
$ARGUMENTSvariable without adequate sanitization. Specifically, in Step 1 ofSKILL.md, if the input does not match a numeric pattern, the raw string is assigned to thePHASEvariable. This variable is subsequently interpolated into double-quoted shell commands such asgrep -A5 "Phase ${PHASE}:" .planning/ROADMAP.md. An attacker can use shell expansion syntax (e.g.,$(command)) within the argument to execute arbitrary code on the host system. - [COMMAND_EXECUTION]: The subagent instructions in
references/phase-research-instructions.mdalso contain shell command templates (e.g.,git add,find,cat) that expand variables derived from potentially untrusted project file names and phase inputs, creating multiple opportunities for command injection or unauthorized file access. - [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. The research agent is explicitly instructed in
references/phase-research-instructions.mdthat information found in*-CONTEXT.mdfiles represents 'locked choices' that 'MUST constrain' its research and behavior. If an attacker or a malicious process can write to these planning files, they can effectively hijack the subagent's instructions to perform unauthorized actions or exfiltrate data via the provided web tools. - Ingestion points: The subagent reads content from
.planning/phases/*/*-CONTEXT.mdand other planning documents. - Boundary markers: The instructions use XML-style tags (e.g.,
<context>) to delineate some fields, but the ingested file content is interpolated into the prompt without escaping or 'ignore embedded instructions' warnings. - Capability inventory: The subagent has access to shell commands (
find,cat,git), web searching (mcp__web_search), and documentation fetching (mcp__context7,WebFetch). - Sanitization: No sanitization or validation of the ingested file content is performed before it is treated as a binding constraint for the agent's logic.
Recommendations
- AI detected serious security threats
Audit Metadata